Nvidia on Wednesday unveiled Agent Knowledge Foundation, a collection of open-source tools that lets enterprise AI agents tap into company information without forcing businesses to shift where that data lives or change who controls access to it. The chip maker introduced the system at its GTC Berlin conference, packaging retrieval models, an ontology library, tabular data models, and data processing tools into its Agent Toolkit. The approach lets companies keep information in existing platforms like Snowflake or Databricks while giving AI agents the context they need to make sense of it all.
Nvidia says its Nemotron Retriever embedding model boosted search relevance by 65% in trials using an open-source OpenSearch configuration, requiring fewer queries and adding almost no delay. SAP's customized version of the reranking tool ran roughly three times quicker, according to Nvidia, while ServiceNow achieved improved reranking performance with nearly 25% lower latency. The company's Auto-Ontology library creates knowledge graphs from corporate databases, mapping raw information to business ideas like customers, products, and orders through a REST API and Model Context Protocol server that agents can query. Nvidia acquired Illumex, an Israeli firm specializing in automated semantic layers and ontologies, for a reported $75 million in February, and bought Kumo for over $400 million in June to add foundation models for relational data.
"I wish it was as easy as connected to my Snowflake system, connected to my email exchange, connected to my knowledge base, and then all of a sudden, magically, the AI understands all the data. But the reality is different," Adel El Hallak, Nvidia's vice president of product for agentic AI, told The New Stack. Large language models excel at predicting tokens but struggle when faced with a single spreadsheet containing multiple tabs, he explained. El Hallak pointed out that knowledge graphs capture concepts and their connections more effectively than retrieval-augmented generation, which simply pulls text chunks based on similarity, adding that "RAG was awesome, but is it good enough?"
The system addresses a core challenge that enterprises face when connecting AI agents to data warehouses: linking to a vendor like Snowflake doesn't automatically give an agent the full business context it needs to operate effectively. Nvidia's approach uses more precise embedding and reranking models that deliver better information to agents upfront, which El Hallak says actually reduces total latency despite adding a small delay to each individual lookup because the agent needs fewer searches overall. The Auto-Ontology library brings consistency across tables when building knowledge graphs, enabling the Kumo models to handle predictive analytics tasks that previously required data teams to construct separate pipelines for each classification, churn analysis, or sentiment task. Nvidia is positioning itself as a partner to data platforms rather than a competitor, with El Hallak noting the company doesn't want to handle access control lists and security features that enterprise partners already provide—Nvidia's own supply chain team uses Agent Knowledge Foundation through Palantir's Foundry platform, Databricks pairs a Kumo anomaly detection model with its Genie Ontology context layer, and Snowflake customers can run Kumo models in Snowflake Notebooks.
El Hallak said Nvidia builds "little weights and models" and publishes them on Hugging Face, which the company agreed to acquire for $12.9 billion in September, leaving data platforms like Snowflake and Databricks in control of the information and permissions while Nvidia supplies the models. That contrasts with SAP's strategy of owning the full stack after closing its acquisition of Prior Labs in July and committing over €1 billion to it across four years, now offering TabPFN-3.5 Plus in SAP AI Core alongside its own RPT models. The division of labor lets Nvidia avoid competing with the data platforms that already store enterprise information and manage the security rules companies have established, instead focusing on the model layer that helps agents interpret what they find. Enterprises seeking flexibility in their AI infrastructure may welcome the option to enhance agent capabilities without consolidating data under a single vendor's control, though the approach depends on partners maintaining robust integration and security standards that Nvidia itself won't manage.